Cryptocurrency Trend Detection Using Hidden Markov Models in Digital Asset Analysis Systems

Cryptocurrency markets are non-stationary with sudden shifts of regimes between market states with different structures (sustained uptrends, long-term downtrends, consolidation, and periods of high volatility periods), which makes linear time series models (ARIMA, exponential smoothing) inappropriate for trend detection as they assume stationarity which is systematically violated by the data. Hidden Markov Models (HMM) offer a principled probabilistic approach to regime identification without requiring stationarity, a natural way to model structural breaks as regime switches, and give rise to posterior state probability distributions at every observation, which can be used for making decisions. This paper uses a four-state Gaussian HMM for 48-month daily returns of five major cryptocurrencies, with 1,461 days of data per asset and 7,305 daily observations per asset (36,525 daily observations pooled across all assets). The four hidden states are identified as S1 (Bull: mean daily return + 1.84%, volatility 2.48%), S2 (Bear: -1.48%, 3.84%), S3 (Sideways: -0.08%, 1.24%), and S4 (High-Volatility: +0.24%, 6.84%). Parameters are estimated using the Baum-Welch Expectation-Maximisation algorithm (convergence criterion is a mean of 48 EM iterations, and a log-likelihood threshold,$\varepsilon=0.001$). Decoding is done using the Viterbi algorithm. The HMM achieves state prediction accuracy of 84.8% on the heldout test set (July-December 2024; 183 days per asset;$\mathrm{n}={9 1 5}$pooled observations) and an AIC of 1,248-substantially below ARIMA$(1,684)$and the naïve baseline (2,048)-while generating trading signals with 78.4% directional accuracy and a mean annualised Sharpe ratio of 1.84 on the test period, compared with 1.24 for ARIMA-signal strategies and 0.84 for buy-and-hold. The state transition matrix reveals that the Bull state has the highest self-persistence probability (0.924), the Bear state shows intermediate persistence (0.848), and the High-Volatility state is the most transient (self-persistence 0.624), consistent with the empirical observation that volatility spikes are short-lived but high-consequence events in cryptocurrency markets.

Hali tarjima qilinmagan